Professional Documents
Culture Documents
By Subhash Chandra Abstract Most economists agree that two major factors affect the shape of the yield curve: investors expectations for future interest rates and certain risk premiums that investors require holding long-term bonds. Because the yield curve can reflect both investors expectations for interest rates and the impact of risk premiums for longer-term bonds, interpreting the yield curve can be complicated. Economists and fixed-income portfolio managers put great effort into trying to understand exactly what forces are driving yields at any given time and at any given point on the yield curve. The way in which these forces simultaneously work to shape the yield curve can be understood. The main objective of this paper is to throw some light on the shape and cause of shapes of yield using Principal Component Analysis. Three factors have been identified, which are almost 99% responsible for the change and shift in the shape of yield curve. A Vector Auto Regressive approach has been applied to those factors, which explains and estimates the shape of yield curve * . 1. Introduction Managing portfolios of financial instruments is in essence managing the tradeoff between risk and return. Optimization is a well suited and frequently used tool to manage this tradeoff. Financial risks arise due to the stochastic nature of some underlying market parameters such as interest rates. So, it is necessary to include stochastic parameters in optimization for portfolio managing, turning portfolio optimization in to stochastic optimization or stochastic programming. A vital part of stochastic programming in portfolio management is scenario generation. Monetary policy makers and observers pay special attention to the shape of the yield curve as an indicator of the impact of current and future monetary policy on the economy. However, drawing inferences from the yield curve is much like reading tea leaves if one does not have the proper tools for yield-curve analysis. The objective is therefore to construct a model capable of capturing the interest rates in order to generate interest rate scenarios. 1.1 What is yield? Yield refers to the annual return on an investment. The yield on a bond is based on both the purchase price of the bond and the interest, or coupon, payments received. There are two ways of looking at bond yields: current yield and yield to maturity.
Current yield is the annual return earned on the price paid for a bond. It is calculated by dividing the bond's annual coupon interest payments by its purchase price. For example, if an investor bought a bond with a coupon rate of 6% at par, and full face value of Rs.1,000, the interest payment over a year would be Rs.60. That would produce a current yield of 6%. When a bond is purchased at full face value, the current yield is the same as the coupon rate. However, if the same bond were purchased at less than face value, or at a discount price, of Rs.900, the current yield would be higher at 6.6%). Yield to maturity reflects the total return an investor receives by holding the bond until it matures. A bonds yield to maturity reflects all of the interest payments from the time of
* It is important to emphasise that the purpose of the model is not to produce superior yield curve predictions, i.e. predictions that in any sense are assumed to out-perform the market and thereby may serve as a basis for tactical investment decisions aimed at outperforming a given benchmark strategy. Rather it is a tool, which supports the investment process related to strategic asset allocation decisions.
270
1.4 What determines the shape of yield curve? Most economists agree that two major factors affect the slope of the yield curve: investors expectations for future interest rates and certain risk premiums that investors require holding long-term bonds. 271
Yield
Because the yield curve can reflect both investors expectations for interest rates and the impact of risk premiums for longer-term bonds, interpreting the yield curve can be complicated. Economists and fixed-income portfolio managers put great effort into trying to understand exactly what forces are driving yields at any given time and at any given point on the yield curve. 2. Available Data Historical data for Indian G-security returns has been taken for all the analysis. Time period chosen is from January 2001 to December 2007. Though daily data was available but here weekly observation has been taken because of non availability of rates for some of the maturity years. Also, data was not available for all the maturity years from January 2001. Some adjustment has been done in F ig.2.1: S hort, Medium & L ong term Y ield that C urve respect. 12 15yL ongTerm Some proxy 4yMediumTerm
10
Y ield (% )
1yS hortTerm
8 6 4 2 J a n01 A pr01 J ul01 O ct01 J a n02 A pr02 J ul02 O ct02 J a n03 A pr03 J ul03 O ct03 J a n04 A pr04 J ul04 O ct04 J a n05 A pr05 J ul05 O ct05 J a n06 A pr06 J ul06 O ct06 J a n07 A pr07 J ul07 O ct07
Dates
observations have been taken. As for example, say, for 3 year maturity observation for one 21st February 2001 is not available. So, as a proxy, observation from nearest maturity year rate has been taken. The data set covers 363 dates with 1 year to 15 year maturity. Because of 272
Yield (%)
273
Table1:
Eigenvalue % of Var. Cum. %
percentage of explanation are shown. Cumulative percentage is also included in Table1. It can be noticed that PC1 (factor1) alone is able to explain more than 98% variation and all the three factors together are explaining 99.6% variations. Annexure A1.1 illustrates
Table2: Coefficients
(2001-2005) PC 1 1yr 0.275 2yr 0.226 3yr 0.223 4yr 0.228 5yr 0.234 6yr 0.243 7yr 0.248 PC 2 -0.714 0.088 0.236 0.321 0.313 0.270 0.195 PC 3 -0.545 -0.356 -0.308 -0.233 -0.196 -0.118 -0.049
274
In Table 2, coefficient of first three most significant factors, when time period chosen was 20012005 have been shown. Column heading in the table are the three most significant factors, whereas row heading are various maturity year. Values inside Table 2 are coefficient of factors, which explain variation in the respective row headings yield. As for example, in case when there is a shock on rates then for 10th row (10yr), PC1 (loading factor1) makes 0.271% variations in 10year bond rate due to the shock; PC2 (loading factor2) explains -0.268% variations in 10year bond rate due to the shock; & PC3 (loading factor3) explains 0.123% variations in 10year bond rate due to the shock. Fig.6 is an example for one particular set of time period (2001-2005). Annexure A1.2 illustrates the plots of first three loading factors for various time periods. It has been observed that first factor (PC1) is almost constant for all the maturity years. However, 2nd factor and 3rd factor are varying with various shapes. Fig.6 shows the three factor loadings corresponding to the three largest principal components in Table 2. The loadings we recognize as the shift, steepness and convexity factors identified by Litterman & Scheinkman. From looking at Fig.6 it can be observed that the first factor forms almost a horizontal line over the whole time period, excluding approximately the first two years. This corresponds to a change of slope for the first two years and a parallel shift for the rest of the maturity horizon. The horizontal line is dominant for the rest of the term structure and hence the factor is recognized as the level factor. The second factor can be interpreted F ig . 6: T hree Mos t S ig nific ant P rinc ipal C omponent as the curvature factor since positive F ac tors (Weekly: J an2001-D ec 2005) F actor1 changes in it cause a decrease in yield 1.0 F actor2 0.8 for bonds with short and long F cator3 0.6 maturities but cause an increase in 0.4 0.2 yield for medium length maturities. 0.0 The third factor is the slope, which 0.2 corresponds to a change of the slope 0.4 0.6 for the whole term structure accounts 0.8 for 0.673% of the total variation. It 1.0 can be seen from the plot that the slope is decreasing as a function of Ma turity maturity which fits the description of a normal yield curve. This is in accordance to the fact that the yield curve the period investigated was for most parts a normal yield cure with marginally diminishing yields.
F ac tor L oading 10yr 11yr 12yr 13yr 14yr
3.3 Effect of Factors on Rates- A unit change of the ith factor causes a change ajt for each maturity t-year rate. Since the factors are independent of each other we may therefore express the total change of the random variable, rt, by
rt = a jt f j
j =1
Where fj is jth factor, k is the number of factors; ajt is the coefficient, identified by the eigenvector analysis, used to approximate the variance of the portfolio. As an example lets see what effect a unit change (_f1 = 1) of the level factor (j = 1) has on the ten year rate (t = 10). From Table2, we have a1,10 = 0.271. so a unit change in factor 1 causes 0.271 275
15yr
1yr
2yr
3yr
4yr
5yr
6yr
7yr
8yr
9yr
Meaning that a 5% ten year rates would become 5.126% if a unit change occurred for all the factors. 4. Choosing the Factors for VAR model The main result from the factor analysis (Principal Component Analysis) was that three factors were to be used to construct the model. But how are the factors recognized in the VAR model? There are two methods for selecting the factors. The 1st method is a naive approach and the 2nd is butterfly method suggested by Christiansen & Lund (2007). The former method is based on taking three positions of the yield curve, a short, medium and long term maturity. The short term rate can be chosen as a proxy for the level factor, the curvature can be chosen as the difference between two yields, a medium maturity yield minus the sort maturity yield. And finally the slope is chosen as two times the medium rate minus the long and short rate. If we note short, medium and long maturity as ys, ym and yl, respectively then the factors can be denoted in the following way level = ys curvature = yl ys slope = 2 ym (ys yl) where we choose the short rate to be the 1 year rate, the medium to be the 4 year and the long to be the 15 year rate (can vary with respect to available data). The 2nd method, butterfly method, is a bit different from nave method. The main difference is that in butterfly approach slope of the yield curve can be chosen differently, namely by using the mechanism of the so called butterfly spread. A butterfly spread is a portfolio which consists of a long position in an intermediate maturity bond (the body of the butterfly) and two short positions of bonds whose maturities straddle the first bond (the wings of the butterfly). Figure below shows a digram of how a butterfly spread looks for a concave (normal) yield curve and the spread s, is given as s = ym (w1*ys + (1 w2)*yl) where the weights w1 and w2 are chosen such that w1ys = w2yl. An example of how the weights are chosen if the maturities are 1, 4 and 15 years would be w1 = (4 1)/(15 1) = 3/14 and weight 2 would become w2 = (154)/(151) = 11/14. The spread shown in the figure is positive and the more concave the yield curve becomes the more positive the spread gets and vice versa. This applies for both normal and inverted yield curves. Equivalently, a negative butterfly spread indicates a convex yield curve.
276
y1,t = c1 + a11 y1,t 1 + a12 y1,t 2 + a13 y1,t 3 + 1,t y 2,t = c 2 + a 21 y 2,t 1 + a 22 y 2,t 2 + a 23 y 2,t 3 + 2,t y 3,t = c3 + a31 y 3,t 1 + a32 y13t 2 + a33 y 3,t 3 + 3,t
In this paper, 2nd method has been applied to select proxy of three factors as described in section 4. Proxy for Factor 1 (level): ys = 1-yr maturity rate Proxy for Factor 2 (curvature): ym = 15-yr maturity rate minus 1-yr maturity rate Proxy for Factor 3 (slope): yl = 4-yr maturity rate minus [ 5.2 Stationary Check For all the three proxies, it has been tested whether they are stationary. It has been found that none of the three proxies are stationary. However, all the three proxies are stationary at their respective 1st difference. A 1st difference for a series, say yt is defined as y t = y t y t 1 . EViews package ahs been used to test the stationarity of the series. Augmented Dicky Fuller (ADF) test has been applied to individual series to test stationarity. Stationary has been decided on the basis of Schwartz Criteria. In Annexure A2, all the stationarity test results have been shown. In Annexure A2.1, it can be seen that neither Level, nor Curvature nor Slope is stationary. For stationary series ADF Test Statistic should be less than critical value. Only proxy for Factor 1 is stationary at 10% level of significance. Annexure A2.2 displays ADF test for 1st order difference series. All the 1st difference series are stationary at all level of significance. Estimation of parameters of VAR model will be done on the basis of these 1st order difference series. 277
5.3 Lag Selection and Criteria For VAR modeling, how many lags are appropriate needs to be identified. EViews package provides facility to identify the lag selection for VAR modeling. Selection of lag has been performed using Schwartz Criteria. 5.4 Estimation of parameters Using three series identified above viz. ( Level ) t , (Curvature) t , ( Slope) t a VAR model has been set up in EViews and the parameter values have been estimated. The results can be seen in Annexure A3. The final equation of the estimation comes out to be as below VAR MODEL: Substituted Coefficient D_CURV = - 0.16766*D_CURV(-1) + 0.10757*D_LEVEL(-1) - 0.131497*D_SLOPE(-1) +0.00167 D_LEVEL = 0.33085*D_CURV(-1) + 0.0974*D_LEVEL(-1) + 0.32341*D_SLOPE(-1) - 0.004674 D_SLOPE =- 0.09689*D_CURV(-1) - 0.07455*D_LEVEL(-1) - 0.16468*D_SLOPE(-1) + 0.000699 In the above table shown, D_LEVEL represents ( Level ) t , D_CURV represents (Curvature) t & D_SLOPE represents ( Slope) t . From these three equations one can identify the equations for actual series very easily. Since above VAR is of order 1, for actual series of factors VAR is of order 2. The actual series, LEVEL (which was representing Factor 1 of our PCA), CURVATURE (which was representing Factor 2 of our PCA) and SLOPE (which was representing Factor 3 of our PCA) can be found from which one can identify the change in rates using equation given section 3.3. 6. Conclusion Using PCA, one can identify the factors which are responsible for changes in yield curve. Modeling VAR is one of the ways to project the future values so that yield to maturity rates can be understood better. Analyzing the VAR process reveled that a process with lag 2 was suitable for modeling the rates, based on the results of information criteria. Investigating the stability of the VAR (2) process reviled that it was stable for the time frame of interest, but using all the data was not necessarily better. Finally, one can apply Vector Error Correction Model (VECM) to take care the shortfall of VAR model. However, along with this other modeling process can also be applied, like ARCH, GARCH or Regime-Switching models. In these ways, investors can prepare their tools, which support the investment process related to strategic asset allocation decisions Bibliography 1. Blaskowitz, O. & Herwartz, H. (2005), Modeling the FIBOR/EURIBOR Swap Term Structure: An Empirical Approach 2. Christiansen, C. & Lund, J. (2007), Revisiting the shape of the yield curve: The effect of interest rate volatility; Working paper 3. Einarsson, A. (2007) Stochastic Scenario Generation for the Term Structure of Interest Rates
278
2001
PC 1 Eigen value 10.755 % of Var. 95.798 Cum. % 95.798 PC 2 0.362 3.226 99.024 PC 3 0.066 0.584 99.608
2001-2002
PC 1 Eigen value 23.734 % of Var. 97.863 Cum. % 97.863 PC 2 0.398 1.641 99.504 PC 3 0.048 0.197 99.701
2001-2004 2001-2003 PC 1 PC 2 Eigen value 37.834 0.310 % of Var. 98.848 0.811 Cum. % 98.848 99.659 PC 3 0.054 0.141 99.800 PC 1 Eigen value 35.947 % of Var. 98.573 Cum. % 98.573 PC 2 0.243 0.665 99.239 PC 3 0.176 0.482 99.721
279
T hree Mos t S ig nific ant P rinc ipal C omponent F ac tors (Weekly: J an2001-D ec 2002)
F acotr1 F actor2 F actor3
F ac to r L o ad in g 1.0 0.8 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0 1y r 2y r 3y r 4y r 5y r 6y r 7y r 8y r 9y r 10y r 11y r 12y r 13y r 14y r
14y r
15yr
Ma turity
Ma turity
T hree Mos t S ig nific ant P rinc ipal C omponent F ac tors (Weekly: J an2001-D ec 2003)
1. 0 0. 8 0. 6 F ac tor L oading 0. 4 0. 2 0. 0 0.2 0.4 0.6 0.8 1.0 10yr 11yr 12yr 13yr 14yr 15yr 1yr 2yr 3yr 4yr 5yr 6yr 7yr 8yr 9yr
T hree Mos t S ig nific ant P rinc ipal C omponent F ac tors (Weekly: J an2001-D ec 2004)
1.0 0.8 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0 1yr 2yr 3yr 4yr 5yr 6yr 7yr 8yr 9yr 10yr 11yr 12yr 13yr
Ma turity
Ma turity
T hree Mos t S ig nific ant P rinc ipal C omponent F ac tors (Weekly: J an2001-D ec 2005)
1. 0 0. 8 0. 6 0. 4 0. 2 0. 0 0.2 0.4 0.6 0.8 1.0 10y r 11y r 12y r 13y r 1y r 2y r 3y r 4y r 5y r 6y r 7y r 8y r 9y r
T hree Mos t S ig nific ant P rinc ipal C omponent F ac tors (Weekly: J an2001-D ec 2006)
14yr
1.0 0.8 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0 10y r 11y r 12y r 13y r
F ac to r L o ad in g
14y r
15y r
Ma turity
Ma turity
T hree Mos t S ig nific ant P rinc ipal C omponent F ac tors (Weekly: J an2001-D ec 2007)
1. 0 0. 8 F ac tor L oading 0. 6 0. 4 0. 2 0. 0 0.2 0.4 0.6 0.8 1.0 1yr 2yr 3yr 4yr 5yr 6yr 7yr 8yr 9yr 10yr 11yr 12yr 13yr 14yr 15yr
Ma turity
280
15y r
1y r
2y r
3y r
4y r
5y r
6y r
7y r
8y r
9y r
15y r
*MacKinnon critical values for rejection of hypothesis of a unit root. Augmented Dickey-Fuller Test Equation Dependent Variable: D(LEVEL) Method: Least Squares Sample(adjusted): 2/05/2001 12/10/2007 Included observations: 358 after adjusting endpoints Variable LEVEL(-1) D(LEVEL(-1)) D(LEVEL(-2)) D(LEVEL(-3)) D(LEVEL(-4)) C R-squared Coefficient Std. Error -0.020808 -0.211934 -0.074128 -0.038336 0.040937 0.128036 0.065358 0.008060 0.052684 0.053825 0.053841 0.052700 0.054108 t-Statistic -2.581681 -4.022744 -1.377192 -0.712019 0.776787 2.366308 Prob. 0.0102 0.0001 0.1693 0.4769 0.4378 0.0185
Adjusted R-squared 0.052082 S.E. of regression 0.199656 Sum squared resid 14.03161
Mean dependent var 0.006927 S.D. dependent var 0.205068 Akaike info criterion 0.367823 Schwarz criterion 0.302787 F-statistic 4.922943 Prob(F-statistic) 0.000228
ADF Test for Curvature -2.920391 1% Critical Value* -3.4515 5% Critical Value -2.8702 10% Critical Value -2.5714
*MacKinnon critical values for rejection of hypothesis of a unit root. Augmented Dickey-Fuller Test Equation Dependent Variable: D(CURVATURE) Method: Least Squares Sample(adjusted): 2/05/2001 12/10/2007 Included observations: 340 Excluded observations: 18 after adjusting endpoints Variable CURVATURE(-1) D(CURVATURE(-1)) D(CURVATURE(-2)) D(CURVATURE(-3)) D(CURVATURE(-4)) C R-squared Coefficient Std. Error -0.067345 -0.228114 -0.026011 -0.038929 0.067013 0.065308 0.099348 281 0.023060 0.056247 0.057706 0.057717 0.055527 0.024741 t-Statistic -2.920391 -4.055587 -0.450749 -0.674481 1.206842 2.639645 Prob. 0.0037 0.0001 0.6525 0.5005 0.2283 0.0087
Adjusted R-squared 0.085865 S.E. of regression 0.190322 Sum squared resid 12.09829
ADF Test for Slope -1.835333 1% Critical Value* -3.4506 5% Critical Value -2.8698 10% Critical Value -2.5712
*MacKinnon critical values for rejection of hypothesis of a unit root. Augmented Dickey-Fuller Test Equation Dependent Variable: D(SLOPE) Method: Least Squares Sample(adjusted): 2/05/2001 12/10/2007 Included observations: 358 after adjusting endpoints Variable SLOPE(-1) D(SLOPE(-1)) D(SLOPE(-2)) D(SLOPE(-3)) D(SLOPE(-4)) C Coefficient Std. Error -0.023005 -0.109507 0.077878 -0.025472 -0.035834 -0.012021 0.012535 0.053465 0.053713 0.053808 0.053496 0.009604 t-Statistic -1.835333 -2.048201 1.449882 -0.473396 -0.669849 -1.251693 Prob. 0.0673 0.0413 0.1480 0.6362 0.5034 0.2115
R-squared 0.034313 Adjusted R-squared 0.020596 S.E. of regression 0.099197 Sum squared resid 3.463671
Mean dependent var 0.002431 S.D. dependent var 0.100234 Akaike info criterion 1.766807 Schwarz criterion 1.701770 F-statistic 2.501477 Prob(F-statistic) 0.030385
ADF Test of 1st Order Difference of Level ADF Test Statistic -8.788308 1% Critical Value* -3.4506 5% Critical Value -2.8698 10% Critical Value -2.5712
*MacKinnon critical values for rejection of hypothesis of a unit root. Augmented Dickey-Fuller Test Equation Dependent Variable: D(D_LEVEL) Method: Least Squares Sample(adjusted): 2/12/2001 12/10/2007 Included observations: 357 after adjusting endpoints Variable D_LEVEL(-1) Coefficient Std. Error -1.271701 0.144704 282 t-Statistic Prob. -8.788308 0.0000
Mean dependent var S.D. dependent var Akaike info criterion Schwarz criterion F-statistic Prob(F-statistic)
ADF Test of 1st Order Difference of Curvature ADF Test Statistic -9.234520 1% Critical Value* -3.4517 5% Critical Value -2.8703 10% Critical Value -2.5714
*MacKinnon critical values for rejection of hypothesis of a unit root. Augmented Dickey-Fuller Test Equation Dependent Variable: D(D_CURV) Method: Least Squares Sample(adjusted): 2/12/2001 12/10/2007 Included observations: 336 Excluded observations: 21 after adjusting endpoints Variable D_CURV(-1) D(D_CURV(-1)) D(D_CURV(-2)) D(D_CURV(-3)) D(D_CURV(-4)) C Coefficient Std. Error -1.439551 0.175199 0.114681 0.041954 0.067202 0.000873 0.155888 0.138866 0.115982 0.089965 0.056201 0.010527 t-Statistic -9.234520 1.261644 0.988780 0.466333 1.195749 0.082955 Prob. 0.0000 0.2080 0.3235 0.6413 0.2327 0.9339
R-squared 0.630998 Adjusted R-squared 0.625407 S.E. of regression 0.192942 Sum squared resid 12.28475
Mean dependent var 0.000506 S.D. dependent var 0.315244 Akaike info criterion 0.435161 Schwarz criterion 0.366999 F-statistic 112.8608 Prob(F-statistic) 0.000000
ADF Test of 1st Order Difference of Slope ADF Test Statistic -8.839290 1% Critical Value* -3.4506 5% Critical Value -2.8698 10% Critical Value -2.5712
*MacKinnon critical values for rejection of hypothesis of a unit root. Augmented Dickey-Fuller Test Equation Dependent Variable: D(D_SLOPE) 283
R-squared 0.573283 Adjusted R-squared 0.567204 S.E. of regression 0.099078 Sum squared resid 3.445593
Mean dependent var 0.000623 S.D. dependent var 0.150604 Akaike info criterion 1.769149 Schwarz criterion 1.703977 F-statistic 94.31187 Prob(F-statistic) 0.000000
Parameter estimation under VAR Sample(adjusted): 1/15/2001 12/10/2007 Included observations: 352 Excluded observations: 9 after adjusting endpoints Standard errors & t-statistics in parentheses D_CURV D_CURV(-1) -0.167660 (0.08885) (-1.88708) 0.107574 (0.09254) (1.16250) -0.131497 (0.11242) (-1.16970) 0.001673 (0.01012) (0.16532) D_LEVEL 0.330859 (0.09078) (3.64466) 0.097427 (0.09455) (1.03043) 0.323407 (0.11486) (2.81556) -0.004674 (0.01034) (-0.45198) 0.086548 0.078673 13.04353 0.193601 10.99075 80.51336 -0.434735 -0.390830 -0.004716 D_SLOPE -0.096887 (0.04623) (-2.09577) -0.074549 (0.04815) (-1.54827) -0.164680 (0.05850) (-2.81528) 0.000699 (0.00527) (0.13269) 0.029830 0.021467 3.382699 0.098592 3.566703 318.0461 -1.784353 -1.740448 0.001026
D_LEVEL(-1)
D_SLOPE(-1)
R-squared 0.078003 Adj. R-squared 0.070055 Sum sq. resids 12.49403 S.E. equation 0.189479 F-statistic 9.813853 Log likelihood 88.08857 Akaike AIC -0.477776 Schwarz SC -0.433871 Mean 0.000369 284
0.201698
0.099668
Determinant Residual 4.10E-06 Covariance Log Likelihood 684.6893 Akaike Information Criteria -3.822098 Schwarz Criteria -3.690384
About the Author: Subhash Chandra Subhash Chandra is Senior Manager, Actuarial in Kotak Mahindra Old Mutual Life Insurance, Mumbai. The views presented in this paper are those of the authors and are not necessarily shared by the organization.
285